Why a Continuity Correction Is Needed
Claim counts take integer values: a frequency can be 17 or 18, but not 17.4. A normal random variable, by contrast, is continuous and assigns probability zero to any single point. When a normal distribution is used to approximate a discrete count, the event boundaries must therefore be adjusted so that the continuous area represents the relevant integer counts. This adjustment is the continuity correction.
The correction is a half-unit shift because adjacent integer counts are one unit apart. It is applied after choosing the count event and before standardizing. The direction of the shift depends on whether the event is a point, a lower tail, an upper tail, or an interval. Applying the right correction does not make a poor normal approximation reliable, so the count model and the suitability of a normal approximation still need to be considered.
Boundary Rules for Count Events
For an inclusive point event, the count \(N=k\) is represented by the half-unit interval centered at \(k\). For a lower or upper tail, the boundary is placed halfway between the last included count and the first excluded count. For an inclusive interval, both endpoints are moved outward by one-half. These rules preserve which integer values belong to the event.
| Discrete event | Continuity-corrected normal event |
|---|---|
| \(N=k\) | \(k-0.5<W<k+0.5\) |
| \(N\leq k\) | \(W<k+0.5\) |
| \(N<k\) | \(W<k-0.5\) |
| \(N\geq k\) | \(W>k-0.5\) |
| \(N>k\) | \(W>k+0.5\) |
| \(a\leq N\leq b\) | \(a-0.5<W<b+0.5\) |
The inequalities at the continuous boundaries can be written as strict or non-strict without changing the probability, since a continuous normal variable has zero probability at any single point. The distinction between \(N\leq k\) and \(N<k\), however, changes the boundary by one full count and must be respected.
If \(W\sim N(\mu,\sigma^2)\), standardize a corrected boundary \(c\) using \(z=(c-\mu)/\sigma\). Write \(\Phi(z)\) for the standard normal cumulative distribution function. For example, a corrected interval from \(L\) to \(U\) has approximate probability \(\Phi((U-\mu)/\sigma)-\Phi((L-\mu)/\sigma)\). Use the count model to obtain \(\mu\) and \(\sigma^2\); the correction changes the event boundaries, not those parameters.
Why the Half-Unit Shift Works
Each possible count can be associated with a unit-width cell centered on that count. For example, the cell for count \(k\) runs from \(k-0.5\) to \(k+0.5\). An inclusive set of counts from \(a\) through \(b\) is therefore represented by adjacent cells running from \(a-0.5\) to \(b+0.5\). Replacing the probability mass of those counts by the normal area over the same span gives the continuity-corrected approximation.
This is a justification of the boundary choice, not an exact identity between the discrete and continuous distributions. The normal curve approximates the count probabilities; it does not turn them into exact normal probabilities. The half-unit shift is useful because it assigns a full unit of continuous width to each integer value being approximated.
Worked Examples
Worked Example: A Binomial Point Probability
A hypothetical portfolio has 64 independent policies. Each policy has probability \(0.5\) of producing a claim during the period, independently of the others. Let \(N\) be the number of claims. Use a normal approximation with a continuity correction to estimate \(P(N=35)\).
Method and model. The count is binomial, \(N\sim\operatorname{Binomial}(64,0.5)\). For a binomial count, \(\mu=np\) and \(\sigma^2=np(1-p)\). Thus \(\mu=64(0.5)=32\), \(\sigma^2=64(0.5)(0.5)=16\), and \(\sigma=4\). We use \(W\sim N(32,16)\) as the normal approximation.
Correct the point event. The count \(N=35\) is represented by \(34.5<W<35.5\). Standardizing both boundaries gives \((34.5-32)/4=0.625\) and \((35.5-32)/4=0.875\). Rounding the \(z\)-values to two decimal places gives \(0.63\) and \(0.88\).
Calculate and check. From a standard normal table, \(\Phi(0.88)=0.8106\) and \(\Phi(0.63)=0.7357\). Therefore the estimate is \(0.8106-0.7357=0.0749\), or about \(7.49\%\). The corrected interval is centered on 35 and includes the single count 35, as required. The common mistake avoided is treating a point probability as the normal probability at exactly 35, which is zero for a continuous variable.
Worked Example: A Poisson Upper Tail
In a hypothetical service unit, the number of claims in a period is modeled as Poisson with parameter \(\lambda=49\). Estimate the probability of at least 58 claims using a normal approximation with a continuity correction.
Method and model. For \(N\sim\operatorname{Poisson}(49)\), the mean is \(\mu=49\) and the variance is \(\sigma^2=49\), so \(\sigma=7\). We approximate \(N\) by \(W\sim N(49,49)\). The event is \(N\geq58\), so the corrected normal event is \(W>57.5\), not \(W>58\).
Standardize and calculate. The standardized boundary is \((57.5-49)/7=8.5/7=1.2143\), which rounds to \(1.21\). Hence \(P(N\geq58)\approx P(Z>1.21)=1-\Phi(1.21)=1-0.8869=0.1131\), where \(Z\sim N(0,1)\). The estimate is \(0.1131\).
Check the boundary. The last count excluded from the event is 57, so the continuous boundary lies halfway between 57 and 58, at 57.5. Using 58 instead would give \(z=(58-49)/7=1.29\) and an estimated upper-tail probability \(1-\Phi(1.29)=0.0985\), materially lower. The correction avoids this common one-count boundary error.
Worked Example: A Negative Binomial Integer Interval
A hypothetical claim-count model describes the number \(N\) of failures observed before the 12th success in independent trials, each with success probability \(p=0.4\). Estimate the probability that \(N\) is between 15 and 22 inclusive.
Method and model. Under this negative binomial parameterization, \(N\) counts failures before \(r\) successes. Its mean is \(r(1-p)/p\), and its variance is \(r(1-p)/p^2\). With \(r=12\) and \(p=0.4\), the mean is \(12(0.6)/0.4=18\), and the variance is \(12(0.6)/(0.4)^2=7.2/0.16=45\). Thus \(\sigma=\sqrt{45}=6.7082\), and we approximate by \(W\sim N(18,45)\).
Correct the interval. The event \(15\leq N\leq22\) becomes \(14.5<W<22.5\). The standardized boundaries are \((14.5-18)/6.7082=-3.5/6.7082=-0.5218\) and \((22.5-18)/6.7082=4.5/6.7082=0.6708\). Rounding to two decimal places gives \(-0.52\) and \(0.67\).
Calculate and check. Using \(\Phi(0.67)=0.7486\) and \(\Phi(-0.52)=0.3015\), the estimate is \(0.7486-0.3015=0.4471\). The result is \(0.4471\), or \(44.71\%\). Both endpoints moved outward, so the corrected continuous interval includes the cells for every integer from 15 through 22. Moving either endpoint inward would omit part of an included count's cell.
Practical Checks and Common Traps
The correction is easiest to apply consistently by identifying the last included integer and the first excluded integer before writing a normal probability. This is especially important for strict inequalities: \(N>k\) starts at \(k+1\), whereas \(N\geq k\) starts at \(k\). Their corrected boundaries are \(k+0.5\) and \(k-0.5\), respectively.
Determine which integer counts are included. Resolve strict versus inclusive inequalities before calculating.
Use the mean and variance of the specified count model, then set \(\sigma=\sqrt{\sigma^2}\).
Use the appropriate point, tail, or interval rule; do not change the mean or variance.
Convert corrected boundaries to \(z\)-scores and use a normal table or calculator. Subtract cumulative probabilities for an interval and take the complement for an upper tail.
A continuity correction improves the alignment between discrete count events and continuous normal areas, but it cannot remedy strong skewness or an unsuitable normal approximation. In particular, a count distribution with substantial skew or a small variance may still be poorly represented. Also, a normal approximation can assign area to negative counts even though a frequency cannot be negative. Use the correction as part of an approximation, not as a guarantee of accuracy.
- Wrong: For \(N=35\), use \(P(W=35)\). A continuous distribution assigns zero probability to one point.
- Correction: Use \(P(34.5<W<35.5)\) to represent the count 35.
- Wrong: For \(N\geq58\), start the normal upper tail at 58.
- Correction: The boundary is 57.5 because 57 is excluded and 58 is included.
- Wrong: For \(15\leq N\leq22\), use boundaries 15 and 22.
- Correction: Expand the interval to 14.5 through 22.5 so the cells for both endpoint counts are included.
Check Your Understanding
For each problem, use a normal approximation with a continuity correction. Round \(z\)-values to two decimal places before reading the standard normal table; report probabilities to four decimal places.
- A hypothetical group of 80 policies has independent claim probabilities of \(0.25\). Estimate the probability of exactly 24 claims. (Answer: 0.0611)
- A claim count is Poisson with parameter \(\lambda=36\). Estimate the probability of at most 30 claims. (Answer: 0.1788)
- A negative binomial count is the number of failures before 8 successes, with success probability \(p=0.5\). Estimate the probability of 10 through 16 failures inclusive. (Answer: 0.3354)
- A hypothetical portfolio has a binomial claim count with \(n=200\) and \(p=0.6\). Estimate the probability of more than 130 claims. (Answer: 0.0643)
- A count is Poisson with parameter \(\lambda=16\). Estimate the probability of either 12 or 13 events. (Answer: 0.1351)